Token

In the sophisticated architecture of modern artificial intelligence, a token represents the fundamental, atomic unit of information that a model processes. Before an algorithm can interpret a sentence, analyze a software script, or recognize objects in an image, the raw input data must be broken down into these discrete, standardized elements. This segmentation is a pivotal step in data preprocessing, transforming unstructured inputs into a numerical format that neural networks can efficiently compute. While humans perceive language as a continuous stream of thoughts or images as seamless visual scenes, computational models require these granular building blocks to perform operations like pattern recognition and semantic analysis.

Token vs. Tokenisation

To grasp the mechanics of machine learning, it is essential to distinguish between the data unit and the process used to create it. This differentiation prevents confusion when designing data pipelines and preparing training material on the Ultralytics Platform.

• Tokenization: This is the algorithmic process (the verb) of splitting raw data into pieces. For text, this might involve using libraries like the Natural Language Toolkit (NLTK) to determine where one unit ends and another begins.
• Token: This is the resulting output (the noun). It is the actual chunk of data—such as a word, a subword, or an image patch—that is eventually mapped to a numerical vector known as an embedding.

Jetons dans différents domaines de l'IA

The nature of a token varies significantly depending on the modality of the data being processed, particularly between textual and visual domains.

Les jetons de texte dans le NLP

In the field of Natural Language Processing (NLP), tokens are the inputs for Large Language Models (LLMs). Early approaches mapped strictly to whole words, but modern architectures utilize subword algorithms like Byte Pair Encoding (BPE). This method allows models to handle rare words by breaking them into meaningful syllables, balancing vocabulary size with semantic coverage. For instance, the word "unhappiness" might be tokenized into "un", "happi", and "ness".

Jetons visuels dans le domaine de la vision par ordinateur

The concept of tokenization has expanded into computer vision with the advent of the Vision Transformer (ViT). Unlike traditional convolutional networks that process pixels in sliding windows, Transformers divide an image into a grid of fixed-size patches (e.g., 16x16 pixels). Each patch is flattened and treated as a distinct visual token. This approach enables the model to use self-attention mechanisms to understand the relationship between distant parts of an image, similar to how Google Research originally applied transformers to text.

Applications concrètes

Les jetons font office de passerelle entre les données humaines et l'intelligence artificielle dans d'innombrables applications.

1. Détection d'objets à vocabulaire ouvert : les modèles avancés tels que YOLO utilisent une approche multimodale dans laquelle les jetons de texte interagissent avec les caractéristiques visuelles. L'utilisateur peut saisir des invites de texte personnalisées (par exemple, « casque bleu »), que le modèle tokenise et compare aux objets présents dans l'image. Cela permet un apprentissage sans tir, permettant la détection d' objets pour lesquels le modèle n'a pas été explicitement entraîné.
2. IA générative : dans les systèmes de génération de texte tels que les chatbots, l'IA fonctionne en prédisant la probabilité du prochain token dans une séquence. En sélectionnant de manière itérative le token suivant le plus probable , le système construit des phrases et des paragraphes cohérents, alimentant des outils allant du service client automatisé aux assistants virtuels.

Python : utilisation de jetons de texte pour la détection

L'extrait de code suivant montre comment le ultralytics package uses text tokens to guide détection d'objets. While the state-of-the-art YOLO26 is recommended for high-speed, fixed-class inference, the YOLO-World architecture uniquely allows users to define classes as text tokens at runtime.

from ultralytics import YOLO

# Load a pre-trained YOLO-World model capable of understanding text tokens
model = YOLO("yolov8s-world.pt")

# Define specific classes; these text strings are tokenized internally
# The model will look specifically for these "tokens" in the visual data
model.set_classes(["bus", "backpack"])

# Run prediction on an image using the defined tokens
results = model.predict("https://ultralytics.com/images/bus.jpg")

# Display the results showing only the tokenized classes
results[0].show()

Understanding tokens is fundamental to navigating the landscape of generative AI and advanced analytics. Whether enabling a chatbot to converse fluently or helping a vision system distinguish between subtle object classes, tokens remain the essential currency of machine intelligence used by frameworks like PyTorch and TensorFlow.